A team of Gemini 4 Argon agents beats Minecraft

We placed four Gemini 4 Argon agents in a safely sandboxed Minecraft survival world and gave them one goal: defeat the Ender Dragon, the boss at the end of the game. These were our general-purpose coding agents, the same ones engineers use for everyday software work, not a custom system built for Minecraft. Operating within a secure environment where researchers monitored a live dashboard and could intervene to pause or stop the run at any time, each agent controlled a player through an open-source library and coordinated with teammates via the game's chat and shared memory. Nobody assigned roles or wrote game logic for them: the agents planned, wrote, and debugged every program themselves.

Twenty-seven hours and 2,783 programs later, they defeated the dragon.

Why does this matter? Beating Minecraft requires maintaining a coherent plan across thousands of dependent steps for over a day. By successfully collaborating, recovering from repeated failure, and pursuing a shared goal without human intervention, these agents proved they can handle the long-term, open-ended problem solving required for complex software engineering and research tasks. We plan to use these learnings and this harness to advance other domains such as coding, science, engineering, and robotics.

The Minecraft Grand Challenge

Finishing Minecraft in a random world requires a long chain of dependent steps: mining basic materials, surviving dangerous alternate dimensions to gather rare items, using math to locate a hidden underground fortress, and finally defeating a boss that constantly heals itself. Each step can take hours, and respawning drops everything you carry. For an agent, the hard part is less any single step than keeping a coherent plan across thousands of them.

Earlier game milestones, from AlphaGo to AlphaStar, came from systems trained at length on a single game, and most Minecraft agents have likewise relied on game-specific training, skill libraries or planners. More recently, community projects have shown that a single language-model agent can reach the dragon. However, these runs relied on heavy constraints and shortcuts: playing on Peaceful mode on a pre-scouted world with a pre-opened portal, or giving a single agent dozens of human-programmed Minecraft tools to execute.

We wanted to ask a different question. Like other large language models, Argon has read plenty about Minecraft on the web. Could a team of general-purpose coding agents take that knowledge and, through a bare programming interface with no game-specific tools, no hints, no human steering, and no pre-assigned roles, organize itself to beat the Ender Dragon in a random survival world from scratch?

Collaborative Coding in Action

The agents never saw a pixel or pressed a key. Each agent acted by writing a short program, such as "walk to these coordinates and mine every iron ore within 16 blocks". They also read the output produced by the program’s execution, including the agent’s updated position, health and inventory, or the error generated. In contrast to previous community efforts, they started from a random world on Easy, with hostile monsters, and the standard Mineflayer library as published, with nothing built on top of it by us. Mineflayer lets a program read a player's position, health, inventory and surroundings, and order basic actions such as walking to a point, digging or placing a block, crafting an item or attacking. Mineflayer does not tell an agent how to play: what to do, and in what order, was up to the agents.

Ten minutes in, the agents had split the work in the chat: one made tools, one gathered food, one dug for diamonds and one built the portal to the game's second dimension. To find the hidden fortress, one agent threw two items that flew toward it from two different places, tracked each flight in code and computed where the two paths crossed. Its estimate landed 43 meters from the fortress, 1.3 kilometers away.

Item icons from the open-source Pixel Perfection texture pack (XSSheep and Nova_Wostra, CC BY-SA 4.0), not Minecraft's textures.

Defeating the dragon was the hardest and most time-consuming challenge in the game, and resulted in the agents respawning 69 times. Between attempts, they wrote down lessons for the next attempt. Eventually, one agent noticed that the crystals were healing the dragon, and worked out that no pickaxe was needed to reach them: stack dirt beside each tower, punch the crystal, dig back down. Thirty-nine minutes after the last crystal fell, the dragon was defeated. The final blow came from a player down to 2.5 of 20 health points.

Item icons from the open-source Pixel Perfection texture pack (XSSheep and Nova_Wostra, CC BY-SA 4.0), not Minecraft's textures.

Along the way, the team kept shared notes of coordinates and lessons, wrote helper functions that all four reused (one appears in 655 programs), recovered each other's items after respawns, and flagged their own mistakes. When a program announced a success that had not happened, its author-agent wrote two minutes later, “My previous chat was an unguarded script branch".

Beyond the Sandbox

Minecraft is the sandbox; the skills are general. Pursuing a goal for more than a day, writing and debugging thousands of programs, recovering from failure, and coordinating as a team are exactly what long-horizon software projects and research demand. They are also what we ask of a good collaborator: an agent that divides the work, keeps shared notes and corrects its own mistakes is one that engineers can trust with real work.

Minecraft is also a safe place to study agent capabilities. The agents' in-game programs could control only their own player: they could neither start processes nor write files, and none used a server command. A fifth agent, the Observer, followed the session as an independent witness to watch for cheating or specification gaming; neither the Observer nor our manual review of a sample of the logs found any. Every program, message and game event was recorded on a live dashboard where researchers could pause or stop the run at any time, and the agents coordinated in plain text and code that anyone could read, line by line. Watching agent teams coordinate in sandboxed environments where every step is visible is one way to understand such systems before they are trusted with real-world tasks.

The limitations of this study are also worth noting. The agents read the world through an API, which tells them exactly which blocks and creatures are around them, and a pathfinding plugin plans their routes around obstacles. They played on Easy. And they failed repeatedly: 605 respawns, and 22% of programs ending in an error.

Verifying the Results

We recorded everything from the run: all 2,783 programs and their outputs, all 229 chat messages, periodic snapshots of the world, and the full server log, where the win is recorded 27 hours and 14 minutes after the first program. Cheat commands were disabled on the server, no program issued a server command or network request, and both the independent Observer agent and our manual audit of the logs confirmed that the team won purely through standard survival gameplay.

The Next Frontier

This study only begins to scratch the surface of using sandbox environments like Minecraft to evaluate how AI teams collaborate. Follow ups include: what happens when agents play directly as humans do from pixels using a virtual keyboard and mouse? How does coordination change when we scale up the size of the team? And what will these agents create when their goal is open-ended as allowed in Minecraft, rather than fixed?

We also plan to use this harness, and the lessons learned from this experiment, to tackle complex real-world challenges, bringing collaborative, long-horizon intelligence to software engineering, scientific research, design, and robotics. The dragon has fallen, but the real work of building dependable AI teammates is just beginning.

Read more about this work in our technical report: http://arxiv.org/abs/2610.22319

This is an independent research experiment. It is not an official Minecraft milestone, and it is not approved by or associated with Mojang or Microsoft. Images of the game use the Pixel Perfection texture pack by XSSheep and Nova_Wostra, licensed under CC BY-SA 4.0, not Minecraft's own textures.